Meteorological Normalisation of PM10 Trends in Switzerland using Random Forest Models

1. J. Sintermann, AWEL PM10 trends in Switzerland using random forest models Stuart K. Grange1,2, David C. Carslaw1,3,andChristophHueglin2 1 Wolfson Atmospheric Chemistry Laboratories, University of York, York, UK 2 Empa, Swiss Federal Laboratories for Materials Science and Technology, Dübendorf, Switzerland 3 Ricardo Energy & Environment, Harwell, Oxfordshire, UK

2. Introduction • Trend analysis of ambient air quality data is a common and important procedure • Air quality trend analysis is complicated because trend can be masked or even driven by changes in meteorology • Controlling for meteorology can be an important component of air quality trend analysis • Various methods have been used for meteorology adjustment/normalisation of air quality observations

3. Objectives • Present a meteorological normalisation technique which uses random forest (RF) to prepare air quality time series for robust trend analysis • Demonstrate the technique using daily PM10 observations across Switzerland between 1997 and 2016 • Details described in Grange et al.ACP (2018) https://www.atmos-chem-phys.net/18/6223/2018/ • Functionality is available in the open source normalweatherrR package (https://github.com/skgrange/normalweatherr)

4. PM10 – observationsfrom 31 sites MapoftheSwiss PM10 and meteorological monitoringsitesincluded in theanalysis

5. Data • Daily PM10observationsweresourcedfromsmonitor* Europe, a databasecontainingdatafromthe European AirBaseandairqualitye-Reporting (AQER) repositories • Surface meteorologicaldataincluding wind speed, wind direction, andtemperatureweresourcedfromthe Integrated Surface Database (ISD) at NOAA https://www.ncdc.noaa.gov/isd • Synopticscalecirculationpatternswereincludedintothemodelsbyusingthe CAP9 Swiss Weather Type Classifications (WTC) • Modelled boundary layer heights were derived from the ERA-Interim data • Clustered HYSPLIT back trajectories were used to represent six distinct air masses • * Grange, S. K., smonitorhttps://github.com/skgrange/smonitor

6. Modelling • RF models using PM10 as the dependent variable for the 31 Swiss monitoring sites were grown • Theexplanatoryvariables were: wind speed, wind direction, temperature, synoptic weather pattern, boundary layer height, air mass cluster, a trend term, a seasonal term, day of the week. • Tuning of RF’s hyper parameters was conducted n=300 80% 20% Conceptualdiagramof a randomforestmodel

7. Meteorologicalnormalisation • The models were used to predict every observed PM10concentration 1000 times (each site separately) • For every prediction, all the explanatory variables, except the trend term were randomly sampled and allocated to a PM10 concentration • All predictions were then aggregated by the mean and this represented “average” meteorological conditions • The Theil-Sen estimator was used to calculate trends

8. Model evaluation • The RF models performed well for most PM10 monitoring sites • R2 values ranged from 54 to 71 % • There was no clear pattern between performance and site type

9. Model evaluation • Importance measures determined the relative contribution of input variables on the prediction of PM10 concentration • Generally, wind speed was the most important variable Importance measures for the 31 RF models. Dots represent the mean and lines represent the interquartile range

10. Trend analysis • Suburban sites decreased at the smallest rates which maybe because these sites are moving to a “more urban” classification • Urban sitesdemonstrated the greatest negative trend magnitudes • Similartowhat has been reported in the past (Barmpadimos et al, ACP, 2011) Meteorologically normalised PM10 trends aggregated by site type

11. Explainingthevariability • RF’s learning processes can be interpreted • Partial dependence plots demonstrate how the explanatory variables are being used for prediction

12. Explainingthevariability • The seasonal component and air temperature partial dependencies were very similar and decision tree methods can handle such collinearity • Very local air masses as well as air mass sourced from Poland and Southern Germany were the most polluted • Boundary layer height: minimum PM10 concentrations when BLH ≈ 1000 metres high but increase again when the BL is over 2000 meters • Weekdays were more polluted than weekends • PM10concentrations were inversely related to wind speed with a non-linear relationship Partial dependence plots for the Zürich-Stampfenbachstrasse RF model

13. Dependencies on BLH suggestspresenceottworegimes conditions which favour convective mixing, transportation, and generation of secondary aerosol e.g. sulphate A poor dispersive environment during wintertime

14. Dependencies on BLH suggestspresenceottworegimes Payerne: Rural (lowaltitude)

15. Conclusions • Controlling for meteorology can be an important component of air quality trend analysis • We presented a meteorological normalisation technique which uses predictive RF models • The technique was tested with daily PM10 observations across Switzerland between 1997 and 2016 • Significantly decreasing trends ranged from -0.09 to -1.16 µg m-3 year-1 • The RF models allow determination of physical and chemical processes influencing ambient PM10 concentrations • - Two regimes resulting in elevated PM10 concentrations were suggested • For a full discussion see Grange et al.ACP (2018) • Functionalitiy available in the normalweatherr R package

16. Thankyou ! References smonitor. https://github.com/skgrange/smonitor. normalweatherr. https://github.com/skgrange/normalweatherr. Grange, S. K., et al. (2018). https://www.atmos-chem-phys.net/18/6223/2018/ Acknowledgements S.K.G thanks Anthony Wild with the provision of the Wild Fund Scholarship.